Transform-based load residual response capability evaluation method and system

Abstract

The invention discloses a load residual response capability evaluation method and system based on Transform, and the method comprises the steps: collecting the operation state of glass factory production equipment in a continuous time step, and forming a high-dimensional state sequence; the method comprises the following steps: performing embedding transformation on a high-dimensional state sequence based on Transform, introducing position coding to construct an input sequence with time sequence perception, performing cross-time-step information fusion by adopting a multi-head self-attention mechanism, and capturing a coupling relationship among equipment, including load, thermal engineering and yield; constructing a glass factory-oriented residual response capability multi-task prediction model and a comprehensive adjustment capability vector; and based on the residual up / down regulation capability and the sustainable response duration, fusing a logic rule and an AI judgment model to judge the glass factory load regulation feasibility. The method integrates the advantages of data driving and logic constraint, provides effective support for the power grid to accurately dispatch the glass industry load, and has good popularization and application values.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a Transformer-based load residual response capability evaluation method and system. Background Art

[0002] Evaluating the residual response capacity of the load in a glass factory helps to understand the flexible adjustment space available in real time for the load, optimize grid dispatching decisions, ensure the stable operation of the power system, reduce production risks, improve energy efficiency, and promote energy conservation, emission reduction and sustainable development. The existing methods for evaluating the load response capacity of glass factories have the following shortcomings: 1) The existing methods lack effective extraction methods for the complex process characteristics of glass factories, especially ignoring the sequence characteristics and inherent correlations of state changes, and fail to deeply reveal the essential relationship between state sequences and load response capacity. 2) The current load forecasting models are mostly simple linear or shallow models, which cannot capture the highly nonlinear characteristics and time series correlations of load changes in the complex process of glass factories, resulting in low residual response capacity prediction accuracy and generalization performance, and poor actual application effect. Due to the above shortcomings, the current power grid is unable to accurately dispatch the load of the glass industry. Summary of the Invention

[0003] In order to solve the above problems, the present invention proposes a Transformer-based load residual response capacity evaluation method and system, which provides effective support for the precise scheduling of glass industry loads by power grids and has good promotion and application value.

[0004] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0005] The Transformer-based load residual response capability evaluation method of the present invention includes the following operations:

[0006] The operating status of the glass factory production equipment is collected at continuous time steps to form a high-dimensional state sequence;

[0007] Based on the Transformer, high-dimensional state sequences are embedded and transformed, position encoding is introduced to construct a time-aware input sequence, and a multi-head self-attention mechanism is used to fuse information across time steps. This captures the coupling relationship between equipment, including load, thermal process, and output, and realizes the extraction of regulation capability features.

[0008] Based on the adjustment capacity features extracted by Transformer, a multi-task prediction model for the remaining response capacity and a comprehensive adjustment capacity vector were constructed for the glass factory. The remaining adjustment capacity (up / down) and sustainable response time were predicted.

[0009] Based on the remaining up / down adjustment capacity and sustainable response time, the feasibility of load regulation in the glass factory is determined by integrating logical rules and AI judgment models.

[0010] A further improvement of the present invention is that the operating status of the glass factory production equipment is collected at continuous time steps to form a high-dimensional state sequence, which specifically includes:

[0011] The data collected includes the power input, operating temperature, material flow, intermediate buffer capacity, and completed historical adjustment records of the melting furnace, forming machine, and annealing furnace, and is uniformly constructed into a state vector. The historical adjustment records include power changes and duration.

[0012] Based on the state vector, a high-dimensional state sequence is constructed on the time window, which is expressed as:

[0013]

[0014] Where: X(t) is the state vector of the glass factory at time t, P m (t) is the furnace power, P f (t) is the power of the molding machine, P a (t) is the annealing furnace power, T m (t) is the furnace temperature, T a (t) is the annealing furnace temperature, S 12 (t) is the intermediate buffer capacity between the melting furnace and the forming section, L m (t) is the output of the melting furnace, L f (t) is the molding output, L a (t) is the annealing yield, Q demand (t) is the current production target, ΔP hist (t) is the historical adjustment power amplitude, τ hist (t) is the duration of historical adjustment, is the constructed glass factory operation state sequence, L is the time window length, X(t-L+1) is the glass factory state vector at time t-L+1, and X(t-L+2) is the glass factory state vector at time t-L+2.

[0015] A further improvement of the present invention is that the expression for extracting the regulation capability feature is:

[0016] E i =W e ·X(t-L+i)+b e ,i=1,2,...,L (3)

[0017] ε (L) =[E1+PE1,E2+PE2,...,E i +PE i ,...,EL +PE L ] (4)

[0018] Q i =ε i W Q ,K i =ε i W K ,V i =ε i W V (5)

[0019] Q=ε (L) W Q ,K=ε (L) W Q ,V=ε (L) W Q (6)

[0020]

[0021] Z (h) =Att (h) (Q (h) ,K (h) ,V (h) ) (8)

[0022] Z=Concat(Z (1) ,...,Z (h) ,...,Z (H) )W O (9)

[0023] Z′=LayerNorm(Z+ε (L) ) (10)

[0024] FFN(x)=max(0,xW1+b1)W2+b2 (11)

[0025] H (L) =LayerNorm(FFN(Z′)+Z′) (12)

[0026] H out =H (L) =[h1,h2,...,h L ] (13)

[0027]

[0028] Where: E i is the embedding vector of the i-th time step, W e is the linear embedding weight matrix, X(t-L+i) is the state vector of the glass factory at time t-L+i, and b eis the bias term, i is the time step, L is the time window length, t is the moment, PE i is the position code of the i-th time step, ε (L) is the input sequence after adding position information, Q i , K i 、V i is the query, key, and value matrix, W Q 、W K 、W V is the attention head weight matrix, ε i is the position information of the state sequence of step i, d k is the dimension of each attention head, Q, K, V are query, key, value, Att(Q,K,V) is the attention, softmax is the function, K Τ is the transpose of the key, Z (h) is the hth attention head output, Z is the output after multi-head splicing, Att (h) (Q (h) ,K (h) ,V (h) ) is the hth attention, Q (h) , K (h) 、V (h) are the h-th query, key, and value respectively, H is the number of attention heads, and W O is the output mapping matrix, LayerNorm(·) is the layer normalization operation, Z′ is the layer normalization operation result, FFN(x), FFN(Z′) are feedforward fully connected networks, x is the input, W1, W2 are the weight matrices of FFN(x), b1, b2 are biases, H out is the output sequence of the Transformer encoder, H (L) is the layer normalization result of the feedforward fully connected network, h t is the state eigenvector, is the final output sequence, h L is the state feature vector of the time window length L.

[0029] A further improvement of the present invention is that the expressions of the constructed residual response capacity multi-task prediction model and the comprehensive regulation capacity vector are:

[0030]

[0031] Where: σ is the activation function, λ1, λ2, λ3 are weight hyperparameters, is the adjustment capacity, that is, the maximum power that the glass factory can still adjust at the current moment. is the downward adjustment capability, that is, the maximum power that can be adjusted downward in the current state, is the maximum power capability available for continued upward regulation, is the maximum power capacity that can still reduce the load, τ remain (t) is the maximum time that the glass factory can continue to adjust under the current state and the set adjustment power level, y t To adjust the capability vector, MLP ↑ (h t ) is the state feature vector h t Mapping to Upward Capacity Nonlinear neural network function, MLP ↓ (h t ) is the state feature vector h t Mapping to a nonlinear neural network function with down-regulation capability, MLP τ (h τ ) is used to predict the longest duration of the adjustment behavior in the current state, W1 ↑ 、 MLP ↑ (h t )’s first and output layer weights, W1 ↓ 、 MLP ↓ (h t )’s first and output layer weights, W1 Τ 、 MLP τ (h τ ), the weights of the first and output layers, is the offset parameter corresponding to the weight, R 3 is the capacity interface vector of the power grid load resource pool, Up-adjust capacity values ​​for real remaining capacity from simulation or historical data, is the maximum power reduction value from actual operation or simulation, In actual observation or simulation, the time value that can be adjusted in the current state is L remain is the sum of the multi-task weighted loss functions, ΔP req To adjust the power request, τ req F is the minimum time requirement for the glass factory to maintain regulation behavior, feas (t) is whether the glass factory has the ability to respond to the current grid regulation request, where a value of 1 indicates feasible and a value of 0 indicates unfeasible. is the maximum time that the glass factory can continuously adjust, and MSE(·) is the mean square error function.

[0032] A further improvement of the present invention is to integrate logical rules and AI judgment models to judge the feasibility of load regulation in glass factories, specifically including:

[0033] Construct a hard-constrained feasibility judgment function to determine whether the current adjustment instruction can be met. The expression is:

[0034]

[0035] Where F′ feas (t) is the feasibility judgment result of the adjustment request based on the hard constraint. It is of Boolean type. The value 1 indicates feasible and the value 0 indicates infeasible. ΔP req To adjust the power request, is the maximum power capability available for continued upward regulation, is the maximum power capacity that can still reduce the load, τ remain (t) is the maximum time that the glass factory can continue to adjust under the current state and the set adjustment power level, τ req The grid dispatcher wants the glass plant to maintain the minimum time requirement for regulation behavior;

[0036] Determine the soft scoring regulation capability margin index, that is, define the response margin factor, including the power margin factor μ P (t), time margin factor μ τ (t), comprehensive margin factor μ total (t), the expression is:

[0037]

[0038] μ total (t) = min(μ P (t),μ τ (t)) (23)

[0039] Build an AI judgment model, the expression is:

[0040]

[0041] F AI (t) = sigmoid(MLP feas (z(t)))∈[0,1] (25)

[0042] Where: z(t) is the input vector, R 5 is a matrix, sigmoid(MLP feas (z(t)))∈[0,1] is the feasibility classifier, F AI (t) is the AI ​​scoring result;

[0043] According to the soft scoring type regulation capacity margin index and AI scoring results, the response level is determined as follows:

[0044]

[0045] Where: A, B, C are adjustment levels.

[0046] The Transformer-based load residual response capability evaluation system of the present invention includes:

[0047] The acquisition module is used to collect the operating status of the glass factory production equipment at continuous time steps to form a high-dimensional state sequence;

[0048] The feature extraction module is used to embed high-dimensional state sequences based on the Transformer. It introduces position encoding to construct a time-aware input sequence. It uses a multi-head self-attention mechanism to fuse information across time steps, capturing the coupling relationship between equipment, including load, thermal performance, and output, to extract regulation capability features.

[0049] The prediction module is used to build a multi-task prediction model for the glass factory's remaining response capacity and a comprehensive regulation capacity vector based on the regulation capacity features extracted by the Transformer. It also predicts the remaining upward / downward regulation capacity and sustainable response time.

[0050] The judgment module is used to determine the feasibility of load regulation in the glass factory based on the remaining up / down adjustment capacity and sustainable response time, integrating logical rules and AI judgment models.

[0051] A further improvement of the present invention is that the specific operations performed by the acquisition module include:

[0052] The data collected includes the power input, operating temperature, material flow, intermediate buffer capacity, and completed historical adjustment records of the melting furnace, forming machine, and annealing furnace, and is uniformly constructed into a state vector. The historical adjustment records include power changes and duration.

[0053] Based on the state vector, a high-dimensional state sequence is constructed on the time window, which is expressed as:

[0054]

[0055] Where: X(t) is the state vector of the glass factory at time t, P m (t) is the furnace power, P f (t) is the power of the molding machine, P a (t) is the annealing furnace power, T m (t) is the furnace temperature, T a (t) is the annealing furnace temperature, S 12 (t) is the intermediate buffer capacity between the melting furnace and the forming section, L m (t) is the output of the melting furnace, L f (t) is the molding output, L a (t) is the annealing yield, Qdemand (t) is the current production target, ΔP hist (t) is the historical adjustment power amplitude, τ hist (t) is the duration of historical adjustment, is the constructed glass factory operation state sequence, L is the time window length, X(t-L+1) is the glass factory state vector at time t-L+1, and X(t-L+2) is the glass factory state vector at time t-L+2.

[0056] A further improvement of the present invention is that the expression for extracting the regulation capability feature is:

[0057] E i =W e ·X(t-L+i)+b e ,i=1,2,...,L (3)

[0058] ε (L) =[E1+PE1,E2+PE2,...,E i +PE i ,...,E L +PE L ] (4)

[0059] Q i =ε i W Q ,K i =ε i W K ,V i =ε i W V (5)

[0060] Q=ε (L) W Q ,K=ε (L) W Q ,V=ε (L) W Q (6)

[0061]

[0062] Z (h) =Att (h) (Q (h) ,K (h) ,V (h) ) (8)

[0063] Z=Concat(Z (1) ,...,Z (h) ,...,Z (H) )W O (9)

[0064] Z′=LayerNorm(Z+ε (L) ) (10)

[0065] FFN(x)=max(0,xW1+b1)W2+b2 (11)

[0066] H (L) =LayerNorm(FFN(Z′)+Z′) (12)

[0067] H out =H (L) =[h1,h2,...,h L ] (13)

[0068]

[0069] Where: E i is the embedding vector of the i-th time step, W e is the linear embedding weight matrix, X(t-L+i) is the state vector of the glass factory at time t-L+i, and b e is the bias term, i is the time step, L is the time window length, t is the moment, PE i is the position code of the i-th time step, ε (L) is the input sequence after adding position information, Q i , K i 、V i is the query, key, and value matrix, W Q 、W K 、W V is the attention head weight matrix, ε i is the position information of the state sequence of step i, d k is the dimension of each attention head, Q, K, V are query, key, value, Att(Q,K,V) is the attention, softmax is the function, K Τ is the transpose of the key, Z (h) is the hth attention head output, Z is the output after multi-head splicing, Att (h) (Q (h) ,K (h) ,V (h) ) is the hth attention, Q (h) , K (h) 、V (h) are the h-th query, key, and value respectively, H is the number of attention heads, and W O is the output mapping matrix, LayerNorm(·) is the layer normalization operation, Z′ is the layer normalization operation result, FFN(x), FFN(Z′) are feedforward fully connected networks, x is the input, W1, W2 are the weight matrices of FFN(x), b1, b2 are biases, H outis the output sequence of the Transformer encoder, H (L) is the layer normalization result of the feedforward fully connected network, h t is the state eigenvector, is the final output sequence, h L is the state feature vector of the time window length L.

[0070] A further improvement of the present invention is that the expressions of the constructed residual response capacity multi-task prediction model and the comprehensive regulation capacity vector are:

[0071]

[0072]

[0073] Where: σ is the activation function, λ1, λ2, λ3 are weight hyperparameters, is the adjustment capacity, that is, the maximum power that the glass factory can still adjust at the current moment. is the downward adjustment capability, that is, the maximum power that can be adjusted downward in the current state, is the maximum power capability available for continued upward regulation, is the maximum power capacity that can still reduce the load, τ remain (t) is the maximum time that the glass factory can continue to adjust under the current state and the set adjustment power level, y t To adjust the capability vector, MLP ↑ (h t ) is the state feature vector h t Mapping to Upward Capacity Nonlinear neural network function, MLP ↓ (h t ) is the state feature vector h t Mapping to a nonlinear neural network function with down-regulation capability, MLP τ (h τ ) is used to predict the longest duration of the adjustment behavior in the current state, W1 ↑ 、 MLP ↑ (h t )’s first and output layer weights, W1 ↓ 、 MLP ↓ (h t )’s first and output layer weights, W1 Τ 、 MLP τ (h τ ), the weights of the first and output layers, is the offset parameter corresponding to the weight, R3 is the capacity interface vector of the power grid load resource pool, Up-adjust capacity values ​​for real remaining capacity from simulation or historical data, is the maximum power reduction value from actual operation or simulation, In actual observation or simulation, the time value that can be adjusted in the current state is L remain is the sum of the multi-task weighted loss functions, ΔP req To adjust the power request, τ req F is the minimum time requirement for the glass factory to maintain regulation behavior, feas (t) is whether the glass factory has the ability to respond to the current grid regulation request, where a value of 1 indicates feasible and a value of 0 indicates unfeasible. is the maximum time that the glass factory can continuously adjust, and MSE(·) is the mean square error function.

[0074] A further improvement of the present invention is that the judgment module performs the following specific operations:

[0075] Construct a hard-constrained feasibility judgment function to determine whether the current adjustment instruction can be met. The expression is:

[0076]

[0077] Where F′ feas (t) is the feasibility judgment result of the adjustment request based on the hard constraint. It is of Boolean type. The value 1 indicates feasible and the value 0 indicates infeasible. ΔP req To adjust the power request, is the maximum power capability available for continued upward regulation, is the maximum power capacity that can still reduce the load, τ remain (t) is the maximum time that the glass factory can continue to adjust under the current state and the set adjustment power level, τ req The grid dispatcher wants the glass plant to maintain the minimum time requirement for regulation behavior;

[0078] Determine the soft scoring regulation capability margin index, that is, define the response margin factor, including the power margin factor μ P (t), time margin factor μ τ (t), comprehensive margin factor μ total (t), the expression is:

[0079]

[0080] μ total (t) = min(μ P (t),μ τ (t)) (23)

[0081] Build an AI judgment model, the expression is:

[0082]

[0083] F AI (t) = sigmoid(MLP feas (z(t)))∈[0,1] (25)

[0084] Where: z(t) is the input vector, R 5 is a matrix, sigmoid(MLP feas (z(t)))∈[0,1] is the feasibility classifier, F AI (t) is the AI ​​scoring result;

[0085] According to the soft scoring type regulation capacity margin index and AI scoring results, the response level is determined as follows:

[0086]

[0087] Where: A, B, C are adjustment levels.

[0088] The present invention comprehensively explores the characteristics of glass factory operating status sequences, improving the dynamics and accuracy of load response capacity assessment. Using Transformer to extract key regulation features, it accurately predicts remaining response capacity and integrates logical rules to assist in determining regulation feasibility, ensuring the safety and practicality of load regulation. By integrating the advantages of data-driven and logical constraints, it provides effective support for the precise scheduling of glass industry loads by power grids, demonstrating its potential for widespread application. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is a schematic diagram of a method flow of an embodiment of the present invention. DETAILED DESCRIPTION

[0090] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0091] like Figure 1 As shown, the Transformer-based load residual response capability evaluation method of this embodiment includes the following operations:

[0092] Step 1: Propose a state sequence model for evaluating the load response capability of a glass factory: collect the operating status of the glass factory's production equipment at continuous time steps to form a high-dimensional state sequence. The data collected include the power input, operating temperature, material flow, intermediate buffer capacity, and completed historical adjustment records of the melting furnace, forming machine, and annealing furnace, and are uniformly constructed into a state vector. The historical adjustment records include power changes and duration. Based on the state vector, a high-dimensional state sequence is constructed on the time window, which comprehensively reflects the historical evolution of the glass factory's load and the dynamic adjustment response trend, providing a time series basis for extracting adjustment capability characteristics. The expression of the state sequence model for evaluating the load response capability of a glass factory is:

[0093]

[0094] Where: X(t) is the state vector of the glass factory at time t, P m (t) is the furnace power, P f (t) is the power of the molding machine, P a (t) is the annealing furnace power, T m (t) is the furnace temperature, T a (t) is the annealing furnace temperature, S 12 (t) is the intermediate buffer capacity between the melting furnace and the forming section, L m (t) is the output of the melting furnace, L f (t) is the molding output, L a (t) is the annealing yield, Q demand (t) is the current production target, ΔP hist (t) is the historical adjustment power amplitude, τ hist (t) is the duration of historical adjustment, is the constructed glass factory operating state sequence, L is the time window length, such as 12 steps represents the past 12 minutes or hours, X(t-L+1) is the glass factory state vector at time t-L+1, and X(t-L+2) is the glass factory state vector at time t-L+2.

[0095] Step 2, extracting regulation capability features based on Transformer: Based on Transformer, embedding transformation is performed on high-dimensional state sequences, and learnable position encoding is introduced to construct an input sequence with time-series awareness. A multi-head self-attention mechanism is used to fuse information across time steps to capture the coupling relationship between equipment including load, thermal process, and output. The multi-head self-attention mechanism of this embodiment strengthens the dependency recognition ability before and after the regulation behavior by weighted combination of historical states. After stacking the Transformer encoder, a comprehensive semantic vector at the current moment is output as a deep representation of the multi-task prediction model of the remaining response capacity, realizing end-to-end extraction of regulation potential and boundary features. Including:

[0096] Perform embedding transformation, as shown in expressions (1) to (2). Expression (1) is to linearly map the state vector of each time step (input embedding), and expression (2) is to add the learnable position encoding (or Sin / Cos encoding) to obtain the final input sequence.

[0097] The multi-head self-attention mechanism is shown in expressions (5) to (9). First, we study single-head attention. Expression (5) calculates the query Q, key K, and value V for each position in the input sequence, i.e., time step i∈{1,...,L}. Expression (6) is the overall matrix form after stacking. Expression (7) calculates the attention output. Second, we study the multi-head mechanism. Expression (8) uses different linear transformations for different attention heads h={1,2,...,H} to obtain the output of each attention head. Expression (9) concatenates all attention heads and maps them back to the original dimension.

[0098] The Transformer encoder output is shown in expressions (10) to (13).

[0099] State compression and output, as shown in expression (14), the semantic vector of the last time step is selected as the representation of the current state, which is used as the input of the multi-task prediction model of the remaining response capacity.

[0100] E i =W e ·X(t-L+i)+b e ,i=1,2,...,L (3)

[0101] ε (L) =[E1+PE1,E2+PE2,...,E i +PE i ,...,E L +PE L ] (4)

[0102] Q i =ε i W Q ,K i =ε i W K ,V i =ε i W V (5)

[0103] Q=ε (L) W Q ,K=ε (L) W Q ,V=ε (L) W Q (6)

[0104]

[0105] Z (h) =Att (h) (Q (h) ,K (h) ,V (h) ) (8)

[0106] Z=Concat(Z (1) ,...,Z (h) ,...,Z (H) )W O (9)

[0107] Z′=LayerNorm(Z+ε (L) ) (10)

[0108] FFN(x)=max(0,xW1+b1)W2+b2 (11)

[0109] H (L) =LayerNorm(FFN(Z′)+Z′) (12)

[0110] H out =H (L) =[h1,h2,...,h L ] (13)

[0111]

[0112] Where: E i is the embedding vector of the i-th time step, W e is the linear embedding weight matrix, X(t-L+i) is the state vector of the glass factory at time t-L+i, and b e is the bias term, i is the time step, L is the time window length, t is the moment, PE i is the position code of the i-th time step, ε (L) is the input sequence after adding position information, Q i , K i 、V i is the query, key, and value matrix, W Q 、W K 、W V is the attention head weight matrix, ε i is the position information of the state sequence of step i, d k is the dimension of each attention head, Q, K, V are query, key, value, Att(Q,K,V) is the attention, softmax is the function, K Τ is the transpose of the key, Z (h) is the hth attention head output, Z is the output after multi-head splicing, Att(h) (Q (h) ,K (h) ,V (h) ) is the hth attention, Q (h) , K (h) 、V (h) are the h-th query, key, and value respectively, H is the number of attention heads, and W O is the output mapping matrix, LayerNorm(·) is the layer normalization operation, Z′ is the layer normalization operation result, FFN(x) is the feedforward fully connected network, x is the input, W1, W2 are the weight matrices of FFN(x), b1, b2 are biases, H out is the output sequence of the Transformer encoder, H (L) is the layer normalization result of the feedforward fully connected network, h t is the state eigenvector, is the final output sequence, h L is the state feature vector of the time window length L.

[0113] Step 3: Based on the regulation capability features extracted by Transformer, a multi-task prediction model for the remaining response capability of the glass factory is constructed. The model includes three output channels, which respectively estimate the maximum power increase and power decrease that the glass factory can currently provide, as well as the sustainable response time at this power level. Each channel is constructed using an independent nonlinear regression network (MLP) with strong nonlinear mapping and generalization capabilities. Through supervised training with historical labels of simulation or field data, the prediction error is minimized and the load regulation margin in the current state is accurately characterized. At the same time, a comprehensive regulation capability vector is constructed for flexible resource aggregation and sorting at the scheduling level. In addition, this module can be extended to applications such as response curve drawing and capacity estimation credible interval output.

[0114] The multi-task prediction model of the remaining response capacity is shown in expression (15). This embodiment predicts the remaining upward adjustment capacity (positive response) separately, that is, and the remaining down-regulation capacity (negative response), i.e. The output unit is kW, reflecting the maximum increase or decrease in the current response of the system.

[0115] The duration prediction can be continuously adjusted as shown in expression (16).

[0116] The comprehensive regulation capability vector construction, as shown in Expression (17), combines the three key outputs into a regulation capability vector, which can be used to support flexibility description, constraint calculation, or sequencing optimization of scheduling.

[0117] The feasibility judgment function of the external adjustment command, as shown in expression (18), can be used as output logic for adjustment command acceptance / rejection judgment or advance planning of response space.

[0118] The model training objective (supervised learning), as shown in expression (19), is to obtain the true labels from the field / simulation dataset A multi-objective mean square error loss function is used.

[0119]

[0120]

[0121] Where: σ is the activation function, L remain is the sum of the multi-task weighted loss functions, which represents the training objective function of the residual responsiveness multi-task prediction model. λ1, λ2, and λ3 are weight hyperparameters that can be optimized through cross-validation to meet the multi-objective requirements of responsiveness evaluation. The power increase capacity is the maximum power that the glass factory can still increase at the current moment (that is, the power capacity that can be further increased without violating the temperature limit, buffer capacity or downstream production requirements). is the downward adjustment capability, that is, the maximum power that can be adjusted downward in the current state, is the maximum power capability available for continued upward regulation, is the maximum power capacity that can still reduce the load, τ remain (t) is the maximum time that the glass factory can continue to adjust under the current state and the set adjustment power level, such as 10% of the rated power, y t MLP is a capacity vector used for flexible load assessment or as a capacity interface vector for power grid load resource pool. ↑ (h t ) is the state feature vector h t Mapping to Upward Capacity A nonlinear neural network function, usually containing 1 to 2 hidden layers, MLP ↓ (h t ) is the state feature vector h t Mapping to a nonlinear neural network function with down-regulation capability, and MLP ↑ (h t ) has the same structure, MLP τ (h τ ) is used to predict the maximum duration of the adjustment behavior in the current state, which is consistent with MLP ↑ (h t ) has the same structure. W1 (·) , are the weights of the first layer and output layer of the MLP, respectively, which are used to control the transformation relationship between the input to the hidden layer and the hidden layer to the output, where: W1 ↑ 、 MLP ↑ (h t )’s first and output layer weights, W1 ↓ 、 MLP ↓ (h t )’s first and output layer weights, W1 Τ 、 MLP τ (h τ ), the weights of the first and output layers, is the offset parameter corresponding to the weight, R 3 is the capacity interface vector of the power grid load resource pool, Up-adjust capacity values ​​for real remaining capacity from simulation or historical data, is the maximum power reduction value from actual operation or simulation, In actual observation or simulation, the time value that can be adjusted in the current state is ΔP req To adjust the power request, positive value is for upward adjustment and negative value is for downward adjustment, τ req F is the minimum time requirement for the glass factory to maintain regulation behavior, feas (t) is whether the glass factory has the ability to respond to the current grid regulation request, where a value of 1 indicates feasible and a value of 0 indicates unfeasible. is the maximum time that the glass factory can continuously adjust, and MSE(·) is the mean square error function, which measures the average quadratic difference between the predicted value and the true label, and is used for the regression class output (power, duration) optimization target.

[0122] Step 4: Based on the remaining up / down regulation capacity and sustainable response time, the feasibility of load regulation of the glass factory is determined by integrating logical rules and AI judgment models to provide regulation strategy decision support for the power grid.

[0123] A hard-constrained feasibility judgment function is constructed to determine whether the current adjustment instruction can be met. Adjustment is allowed only when the adjustment request direction matches, the amplitude does not exceed the capacity, and the duration does not exceed the safety boundary.

[0124]

[0125] Where F′ feas (t) is the feasibility judgment result of the adjustment request based on the hard constraint. It is of Boolean type. The value 1 indicates feasible and the value 0 indicates infeasible. ΔP reqTo adjust the power request, is the maximum power capability available for continued upward regulation, is the maximum power capacity that can still reduce the load, τ remain (t) is the maximum time that the glass factory can continue to adjust under the current state and the set adjustment power level, τ req The grid dispatcher wants the glass plant to maintain the minimum time requirement for regulation behavior;

[0126] Determine the soft scoring regulation capability margin index, that is, define the response margin factor, including the power margin factor μ P (t), time margin factor μ τ (t), comprehensive margin factor μ total (t). It can be used in scenarios such as multi-device response sorting and flexible regional scheduling and allocation.

[0127]

[0128] μ total (t) = min(μ P (t),μ τ (t)) (23)

[0129] Power margin factor μ P (t) is dimensionless and reflects the redundancy of the power regulation of the glass factory. The larger the value, the more "headroom" there is in the current response request power. The time margin factor μ τ (t) is dimensionless and reflects the time redundancy of the glass factory for the request duration. The comprehensive margin factor μ total (t) represents the minimum margin of the current regulation request in terms of power amplitude and duration.

[0130] Construct an AI judgment model. When process constraints are complex and the influence of adjustment history is strong, a lightweight neural network model can be used for judgment. Expression (24) defines the input vector, and Expression (25) constructs a feasibility classifier. The closer the output value is to 1, the more feasible the adjustment is; the closer the output value is to 0, the greater the probability that the adjustment violates the thermal / process / cache boundaries.

[0131]

[0132] F AI (t) = sigmoid(MLP feas (z(t)))∈[0,1] (25)

[0133] Where: z(t) is the input vector, R 5 is a matrix, sigmoid(MLP feas (z(t)))∈[0,1] is the feasibility classifier, FAI (t) is the AI ​​scoring result.

[0134] The response level is determined based on the soft-scoring regulatory capacity margin index and AI scoring results.

[0135]

[0136] Where: A, B, C are adjustment levels.

[0137] The present invention integrates the advantages of data-driven and logical constraints, provides effective support for the precise dispatching of glass industry loads by power grids, and has good promotion and application value.

[0138] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless defined similarly as herein, will not be interpreted in an idealized or overly formal sense.

[0139] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. Transformer-based load residual response capacity assessment method, characterized by: The following operations are included: The operating status of the glass factory production equipment is collected at continuous time steps to form a high-dimensional state sequence; Based on the Transformer, high-dimensional state sequences are embedded and transformed, position encoding is introduced to construct a time-aware input sequence, and a multi-head self-attention mechanism is used to fuse information across time steps. This captures the coupling relationship between equipment, including load, thermal process, and output, and realizes the extraction of regulation capability features. Based on the adjustment capacity features extracted by Transformer, a multi-task prediction model for the remaining response capacity and a comprehensive adjustment capacity vector were constructed for the glass factory. The remaining adjustment capacity (up / down) and sustainable response time were predicted. Based on the remaining up / down adjustment capacity and sustainable response time, the feasibility of load regulation in the glass factory is determined by integrating logical rules and AI judgment models.

2. The Transformer-based load residual response capability assessment method according to claim 1, characterized in that: The operation status of the glass factory production equipment is collected at continuous time steps to form a high-dimensional state sequence, specifically including: The data collected includes the power input, operating temperature, material flow, intermediate buffer capacity, and completed historical adjustment records of the melting furnace, forming machine, and annealing furnace, and is uniformly constructed into a state vector. The historical adjustment records include power changes and duration. Based on the state vector, a high-dimensional state sequence is constructed on the time window, which is expressed as: χ t (L) =[X(t-L+1),X(t-L+2),...,X(t)] (2) Where: X(t) is the state vector of the glass factory at time t, P m (t) is the furnace power, P f (t) is the power of the molding machine, P a (t) is the annealing furnace power, T m (t) is the furnace temperature, T a (t) is the annealing furnace temperature, S 12 (t) is the intermediate buffer capacity between the melting furnace and the forming section, L m (t) is the output of the melting furnace, L f (t) is the molding output, L a (t) is the annealing yield, Q demand (t) is the current production target, ΔP hist (t) is the historical adjustment power amplitude, τ hist (t) is the duration of historical adjustment, χ t (L) is the constructed glass factory operation state sequence, L is the time window length, X(t-L+1) is the glass factory state vector at time t-L+1, and X(t-L+2) is the glass factory state vector at time t-L+2.

3. The Transformer-based load residual response capability assessment method according to claim 1, characterized in that: The expression for extracting the adjustment capability feature is: E i =W e ·X(t-L+i)+b e ,i=1,2,...,L (3) ε (L) =[E1+OR1,E2+OR2,...,E i +PE i ,...,BY L +PE L ] (4) Q i =e i W Q ,K i =e i W K ,V i =e i W V (5) Q=e (L) W Q ,K=e (L) W Q ,V=ε (L) W Q (6) Z (h) =To (h) (Q (h) ,K (h) ,V (h) ) (8) Z=Concat(Z (1) ,...,WITH (h) ,...,WITH (H) )IN O (9) Z′=LayerNorm(Z+ε (L) ) (10) FFN(x)=max(0,xW1+b1)W2+b2 (11) H (L) =LayerNorm(FFN(Z′)+Z′) (12) H out =H (L) =[h1,h2,...,h L ] (13) Where: E i is the embedding vector of the i-th time step, W e is the linear embedding weight matrix, X(t-L+i) is the state vector of the glass factory at time t-L+i, and b e is the bias term, i is the time step, L is the time window length, t is the moment, PE i is the position code of the i-th time step, ε (L) is the input sequence after adding position information, Q i , K i 、V i is the query, key, and value matrix, W Q 、W K 、W V is the attention head weight matrix, ε i is the position information of the state sequence of step i, d k is the dimension of each attention head, Q, K, V are query, key, value, Att(Q,K,V) is the attention, softmax is the function, K Τ is the transpose of the key, Z (h) is the hth attention head output, Z is the output after multi-head splicing, Att (h) (Q (h) ,K (h) ,V (h) ) is the hth attention, Q (h) , K (h) 、V (h) are the h-th query, key, and value respectively, H is the number of attention heads, and W O is the output mapping matrix, LayerNorm(·) is the layer normalization operation, Z′ is the layer normalization operation result, FFN(x), FFN(Z′) are feedforward fully connected networks, x is the input, W1, W2 are the weight matrices of FFN(x), b1, b2 are biases, H out is the output sequence of the Transformer encoder, H (L) is the layer normalization result of the feedforward fully connected network, h t is the state eigenvector, is the final output sequence, h L is the state feature vector of the time window length L.

4. The Transformer-based load residual response capability assessment method according to claim 1, characterized in that: The expressions of the constructed residual response capacity multi-task prediction model and comprehensive regulation capacity vector are as follows: Where: σ is the activation function, λ1, λ2, λ3 are weight hyperparameters, is the adjustment capacity, that is, the maximum power that the glass factory can still adjust at the current moment. is the downward adjustment capability, that is, the maximum power that can be adjusted downward in the current state, is the maximum power capability available for continued upward regulation, is the maximum power capacity that can still reduce the load, τ remain (t) is the maximum time that the glass factory can continue to adjust under the current state and the set adjustment power level, y t To adjust the capability vector, MLP ↑ (h t ) is the state feature vector h t Mapping to Upward Capacity Nonlinear neural network function, MLP ↓ (h t ) is the state feature vector h t Mapping to a nonlinear neural network function with down-regulation capability, MLP τ (h τ ) is used to predict the longest duration of the adjustment behavior in the current state, W1 ↑ 、W2 ↑ MLP ↑ (h t )’s first and output layer weights, W1 ↓ 、W2 ↓ MLP ↓ (h t )’s first and output layer weights, W1 Τ 、W2 Τ MLP τ (h τ ), the weights of the first and output layers, is the offset parameter corresponding to the weight, R 3 is the capacity interface vector of the power grid load resource pool, Up-adjust capacity values ​​for real remaining capacity from simulation or historical data, is the maximum power reduction value from actual operation or simulation, In actual observation or simulation, the time value that can be adjusted in the current state is L remain is the sum of the multi-task weighted loss functions, ΔP req To adjust the power request, τ req F is the minimum time requirement for the glass factory to maintain regulation behavior, feas (t) is whether the glass factory has the ability to respond to the current grid regulation request, where a value of 1 indicates feasible and a value of 0 indicates unfeasible. is the maximum time that the glass factory can continuously adjust, and MSE(·) is the mean square error function.

5. The Transformer-based load residual response capability assessment method according to claim 1, characterized in that: The fusion logic rules and AI judgment model are used to judge the feasibility of load regulation in the glass factory, specifically including: Construct a hard-constrained feasibility judgment function to determine whether the current adjustment instruction can be met. The expression is: Where, F f ' eas (t) is the feasibility judgment result of the adjustment request based on the hard constraint. It is of Boolean type. The value 1 indicates feasible and the value 0 indicates infeasible. ΔP req To adjust the power request, is the maximum power capability available for continued upward regulation, is the maximum power capacity that can still reduce the load, τ remain (t) is the maximum time that the glass factory can continue to adjust under the current state and the set adjustment power level, τ req The grid dispatcher wants the glass plant to maintain the minimum time requirement for regulation behavior; Determine the soft scoring regulation capability margin index, that is, define the response margin factor, including the power margin factor μ P (t), time margin factor μ τ (t), comprehensive margin factor μ total (t), the expression is: m total (t)=min(μ P (t),μ τ (t)) (23) Build an AI judgment model, the expression is: F AI (t)=sigmoid(MLP feas (z(t)))∈[0,1] (25) Where: z(t) is the input vector, R 5 is a matrix, sigmoid(MLP feas (z(t)))∈[0,1] is the feasibility classifier, F AI (t) is the AI ​​scoring result; According to the soft scoring type regulation capacity margin index and AI scoring results, the response level is determined as follows: Where: A, B, C are adjustment levels.

6. Transformer-based load excess response capability assessment system, characterized by: include: The acquisition module is used to collect the operating status of the glass factory production equipment at continuous time steps to form a high-dimensional state sequence; The feature extraction module is used to embed high-dimensional state sequences based on the Transformer. It introduces position encoding to construct a time-aware input sequence. It uses a multi-head self-attention mechanism to fuse information across time steps, capturing the coupling relationship between equipment, including load, thermal performance, and output, to extract regulation capability features. The prediction module is used to build a multi-task prediction model for the glass factory's remaining response capacity and a comprehensive regulation capacity vector based on the regulation capacity features extracted by the Transformer. It also predicts the remaining up / down regulation capacity and sustainable response time. The judgment module is used to determine the feasibility of load regulation in the glass factory based on the remaining up / down adjustment capacity and sustainable response time, integrating logical rules and AI judgment models.

7. The Transformer-based load excess response capability assessment system according to claim 6, characterized in that: The specific operations performed by the acquisition module include: The data collected includes the power input, operating temperature, material flow, intermediate buffer capacity, and completed historical adjustment records of the melting furnace, forming machine, and annealing furnace, and is uniformly constructed into a state vector. The historical adjustment records include power changes and duration. Based on the state vector, a high-dimensional state sequence is constructed on the time window, which is expressed as: χ t (L) =[X(t-L+1),X(t-L+2),...,X(t)] (2) Where: X(t) is the state vector of the glass factory at time t, P m (t) is the furnace power, P f (t) is the power of the molding machine, P a (t) is the annealing furnace power, T m (t) is the furnace temperature, T a (t) is the annealing furnace temperature, S 12 (t) is the intermediate buffer capacity between the melting furnace and the forming section, L m (t) is the output of the melting furnace, L f (t) is the molding output, L a (t) is the annealing yield, Q demand (t) is the current production target, ΔP hist (t) is the historical adjustment power amplitude, τ hist (t) is the duration of historical adjustment, χ t (L) is the constructed glass factory operation state sequence, L is the time window length, X(t-L+1) is the glass factory state vector at time t-L+1, and X(t-L+2) is the glass factory state vector at time t-L+2.

8. The Transformer-based load excess response capability assessment system according to claim 6, characterized in that: The expression for extracting the adjustment capability feature is: E i =W e ·X(t-L+i)+b e ,i=1,2,...,L (3) ε (L) =[E1+OR1,E2+OR2,...,E i +PE i ,...,BY L +PE L ] (4) Q i =e i W Q ,K i =e i W K ,V i =e i W V (5) Q=e (L) W Q ,K=e (L) W Q ,V=ε (L) W Q (6) Z (h) =To (h) (Q (h) ,K (h) ,V (h) ) (8) Z=Concat(Z (1) ,...,WITH (h) ,...,WITH (H) )IN O (9) Z′=LayerNorm(Z+ε (L) ) (10) FFN(x)=max(0,xW1+b1)W2+b2 (11) H (L) =LayerNorm(FFN(Z′)+Z′) (12) H out =H (L) =[h1,h2,...,h L ] (13) Where: E i is the embedding vector of the i-th time step, W e is the linear embedding weight matrix, X(t-L+i) is the state vector of the glass factory at time t-L+i, and b e is the bias term, i is the time step, L is the time window length, t is the moment, PE i is the position code of the i-th time step, ε (L) is the input sequence after adding position information, Q i , K i 、V i is the query, key, and value matrix, W Q 、W K 、W V is the attention head weight matrix, ε i is the position information of the state sequence of step i, d k is the dimension of each attention head, Q, K, V are query, key, value, Att(Q,K,V) is the attention, softmax is the function, K Τ is the transpose of the key, Z (h) is the hth attention head output, Z is the output after multi-head splicing, Att (h) (Q (h) ,K (h) ,V (h) ) is the hth attention, Q (h) , K (h) 、V (h) are the h-th query, key, and value respectively, H is the number of attention heads, and W O is the output mapping matrix, LayerNorm(·) is the layer normalization operation, Z′ is the layer normalization operation result, FFN(x), FFN(Z′) are feedforward fully connected networks, x is the input, W1, W2 are the weight matrices of FFN(x), b1, b2 are biases, H out is the output sequence of the Transformer encoder, H (L) is the layer normalization result of the feedforward fully connected network, h t is the state eigenvector, is the final output sequence, h L is the state feature vector of the time window length L.

9. The Transformer-based load excess response capability assessment system according to claim 6, characterized in that: The expressions of the constructed residual response capacity multi-task prediction model and comprehensive regulation capacity vector are as follows: Where: σ is the activation function, λ1, λ2, λ3 are weight hyperparameters, is the adjustment capacity, that is, the maximum power that the glass factory can still adjust at the current moment. is the downward adjustment capability, that is, the maximum power that can be adjusted downward in the current state, is the maximum power capability available for continued upward regulation, is the maximum power capacity that can still reduce the load, τ remain (t) is the maximum time that the glass factory can continue to adjust under the current state and the set adjustment power level, y t To adjust the capability vector, MLP ↑ (h t ) is the state feature vector h t Mapping to Upward Capacity Nonlinear neural network function, MLP ↓ (h t ) is the state feature vector h t Mapping to a nonlinear neural network function with down-regulation capability, MLP τ (h τ ) is used to predict the maximum duration that the adjustment behavior can last in the current state, MLP ↑ (h t ), the weights of the first and output layers, MLP ↓ (h t ), the weights of the first and output layers, MLP τ (h τ ), the weights of the first and output layers, is the offset parameter corresponding to the weight, R 3 is the capacity interface vector of the power grid load resource pool, Up-adjust capacity values ​​for real remaining capacity from simulation or historical data, is the maximum power reduction value from actual operation or simulation, In actual observation or simulation, the time value that can be adjusted in the current state is L remain is the sum of the multi-task weighted loss functions, ΔP req To adjust the power request, τ req F is the minimum time requirement for the glass factory to maintain regulation behavior, feas (t) is whether the glass factory has the ability to respond to the current grid regulation request, where a value of 1 indicates feasible and a value of 0 indicates unfeasible. is the maximum time that the glass factory can continuously adjust, and MSE(·) is the mean square error function.

10. The Transformer-based load excess response capability assessment system according to claim 6, characterized in that: The specific operations performed by the judgment module include: Construct a hard-constrained feasibility judgment function to determine whether the current adjustment instruction can be met. The expression is: Where, F f ' eas (t) is the feasibility judgment result of the adjustment request based on the hard constraint. It is of Boolean type. The value 1 indicates feasible and the value 0 indicates infeasible. ΔP req To adjust the power request, is the maximum power capability available for continued upward regulation, is the maximum power capacity that can still reduce the load, τ remain (t) is the maximum time that the glass factory can continue to adjust under the current state and the set adjustment power level, τ req The grid dispatcher wants the glass plant to maintain the minimum time requirement for regulation behavior; Determine the soft scoring regulation capability margin index, that is, define the response margin factor, including the power margin factor μ P (t), time margin factor μ τ (t), comprehensive margin factor μ total (t), the expression is: Build an AI judgment model, the expression is: F AI (t)=sigmoid(MLP feas (z(t)))∈[0,1] (25) Where: z(t) is the input vector, R 5 is a matrix, sigmoid(MLP feas (z(t)))∈[0,1] is the feasibility classifier, F AI (t) is the AI ​​scoring result; According to the soft scoring type regulation capacity margin index and AI scoring results, the response level is determined as follows: Where: A, B, C are adjustment levels.

Images